Idan Segev Interdisciplinary Center for Neural Computation Hebrew University Thanks to: Miki London Galit Fuhrman Adi Shraibman Elad Schneidman What does the synapse tell the axon?

Outline Introduction Questions in my group A brief history of the synapse what does “synaptic efficacy” mean? Complications with “synaptic efficacy” Information theory (I.T.) and synaptic efficacy Basic definitions (entropy, compression & mutual information) The “noisy input-output” model Preliminary Results “Synaptic efficacy” in the context of I.T. In simple neuron models In passive dendritic structures In excitable dendrites Conclusions Future questions

1. Neuronal “noise” and input-output properties of neurons (Elad Schneidmann, Miki London) Ion-channels, synaptic noise and AP reliability Optimization of information transmission with noise 2. Nonlinear cable theory (Claude Meunier) Threshold conditions for excitation in excitable dendrites Active propagation in excitable trees 3. “Learning rules” for ion channels and synapses. How to build a “H&H” axon? How to “read” synaptic plasticity? 4. The synapse: “what does it say”? (Miki London, Galit Fuhrman) Could dynamic synapses encode the timing of the pre-synaptic spikes? “Synaptic efficacy” - what does it mean? Research focus in my group

The “father” of the  Sir Charles Scott Sherrington Syndesm (“connection”) - Sherrington Synapsis (“Clasp”) - Verrall (Greek scholar/Cambridge) “Each synapsis offers an opportunity for a change in the character of nervous impulses, that the impulse as it passes over from the terminal arborescence of an axon into the dendrite of another cell, starts in that dendrite an impulse having character different from its own” Forster and Sherrington, 1897

Whitney Museum Presents: Synapsis Shuffle, a New Masterwork by Robert Rauschenberg Robert Rauschenberg has organized a hodgepodge group of famous names; from the highbrow (Robert Hughes, Chuck Close) to the lowbrow (Martha Stewart, Michael Ovitz) around the not-especially radical idea that anyone can create a Rauschenberg. Each participant chose an image (by lottery) from a total of 52 Rauschenberg transfer photographs, and then created a composition. “Blown Synapses” The result is bland, homogeneous work on an unnecessarily large scale. Perhaps if the project's parameters had been more narrowly defined?say, if each participant were allotted the same five images?these works would offer more insight into the minds of their composers. As it is, Rauschenberg's shuffle dulls the synapses. Karen Rosenberg ”

Motivation: Single synapse matters 400 ext. (10/sec) 100 inh. (65/sec) Mainen & Sejnowki model

Motivation: Single synapse matters 200 sec simulation (10 spikes/sec)

Motivation: Single synapse matters

“Synaptic efficacy” Artificial Neural Networks - synaptic efficacy reduced to a single number, W ij (J ij ) Biophysics - Utilizing the (average) properties of the PSP (peak; rise-time; area, charge …) Cross-Correlation - Relating the pre-synaptic input to the post-synaptic output (the firing probability). But how to interpret the shape of the cross-correlation?

Complications with “synaptic efficacy”: PSP have different shape indices: Who is more “effective” and by how much? EPSP peak is equal but the rise time is different EPSP area (charge) is equal but the peak is different

Complications with “synaptic efficacy”: Synapses are dynamic Facilitating Depressing

Complications with “synaptic efficacy”: The synapse: a voice in the crowd synaptic effect depends on the context (and the synapse itself is probabilistic) L.J. Borg-Graham, C. Monier & Y. Frengac Spontaneous in vivo voltage fluctuations in a neuron from the cat visual cortex

A new definition for “Synaptic efficacy” “Neuron” Output Input Noise Background Activity Input Output Mutual Information Mutual information: what does the synaptic input tell us about the spike output? “Synaptic efficacy”: The mutual information between the input and the output ?

Entropy Known Synaptic Input The Mutual Information (MI) is the extra bits saved in encoding the output by knowing the input Computing the mutual information (Compression, Entropy and Mutual Information) Information in the input? Output Spike train Compressed Spike train output Compressed output Spike train given the input Mutual Information Compression Information estimation We use the CTW compression algorithm (best known today)

Mutual information in a Simple I&F model (effect of potentiation) Threshold Isolated synapse background Background synapse x5x5 Output spike train

Which of the EPSP parameters affects the MI? Fixed peak Fixed charge the MI corresponds best to the EPSP peak

Why the MI corresponds best to EPSP peak? Sharp EPSP Broad EPSP Less spikes, More accurate More spikes, Less accurate Input

M.I (“synaptic efficacy”) in realistic models: Passive Cable with (linear) synapses +H&H axon

(Cable with linear synapses) MI (synaptic efficacy) of distal synapses scales with EPSP peak Proximal Distal

MI with Active dendritic currents proximal distal intermediate The active boosting affects both input synapse but also the “background noise” (i) Proximal synapse transmits less information compared to passive case (“noise” is larger and proximal EPSP is almost passive) (ii) Distal synapse is relatively more boosted due to large local input impedance. (iii) Intermediate synapse is boosted as much as the noise does; so it does not transmit more information in the active case.

Conclusions The mutual information measure provides a functional link between the synaptic input and the spike output. Hence, the M.I could be interpreted as “synaptic efficacy”. “Synaptic efficacy” depends on the context within which the synapse operates. The EPSP peak (rather than its area) corresponds most closely to the mutual information. Active dendritic currents affect both the “background noise”and the input synapse. The relative effect of this noise on the “efficacy of the synaptic input” depends on the location of the input. Typically, distal synapses tend to be relatively more boosted.

Future Questions Natural Generalizations for charaterizing “synaptic efficacy” *MI (efficacy) of Inhibitory synapses *Depressing, facilitating and probabilistic synapses *Dependence on input structure (regular input; bursting input) *Dependence on the Context (correlated background) *Dependence on dendrtic excitability (I h, I A, I Ca,.., ) *Dependence on # of and site of connection “synaptic efficacy” for many pre-synaptic inputs “Selfish” or Cooperative strategies for maximizing information transfer (each synapse may want to increase its own EPSP peak, but others do too)

Effect of bin size Sharp Wide Sharp Control x3 x5

12,000 Na channels 3,600 K channels 200  m 2